Research on the Improved Back Propagation
Neural Network for the Aquaculture Water
Quality Prediction
Ding Jin-ting and Zang Ze-lin School of Information & Electrical Engineering, Zhejiang University City College, Hangzhou, China
Email: [email protected], [email protected]
Abstract—In order to overcome the slow convergence speed,
easy to fall into the shock and generalization ability is not
strong and so on of the traditional BP neural network, the
adaptive variable step size BP neural network learning
algorithm based on the fuzzy method has been developed,
which can reduce the learning time of BP neural network,
improve the convergence efficiency and network stability.
According to the water quality monitoring data from one of
Penaeus vannamei breeding base in Hangzhou, a
mathematical model of multi-layer feed forward network
was established to predict and evaluate the quality of the
aquaculture water. The topology of the model is 40-14-4,
that is, the temperature (T), pH value, Dissolved Oxygen
(DO), and the Redox Potential (ORP) are the input variables
in n=10 consecutive time units, the number of hidden layer
nodes is 14, and the output layer is 4. The results show that
the improved BP neural network based on fuzzy method has
the characteristics of fast convergence, high accuracy and
good stability. It provides a new method for the prediction
and evaluation of water quality in aquaculture.
Index Terms—artificial neural network, BP network, water
quality prediction, adaptive variable step size algorithm
I. INTRODUCTION
With the development of aquaculture technology and
the expansion of the scale of farming, how to reduce the
risk of breeding has become an important issue. How to
establish a sensor monitoring of aquatic products life
became more and more popular choice [1]-[3]. However,
aquaculture water has large volume, the nonlinear
variation characteristics, found that the water beyond the
alarm boundary in the take remedial measures is often not
timely to recoup their losses, so it is necessary to forecast
and alarm of water quality changes.
In order to solve the problem of water quality
prediction and alarm, a lot of research have been carried
out. Liu Shuang-yin [4] gives the aquaculture water
quality pH value forecasting model by the principal
component analysis and least squares support vector
machine method, but pH value output obviously cannot
meet the needs of aquaculture. Hu Hai-qing [5] has
settled the new modeling and forecasting on river water
pH and turbidity based on LM-BP neural network, which
Manuscript received March 22, 2016; revised July 28, 2016.
was focused on using a time parameter values to predict
next time parameter value, does not consider the change
tendency of the parameters extraction and result in bad
prediction accuracy. Yuan Qi [6] employed the adaptive
BP neural network algorithm to model and predict the pH
value and turbidity of the culture water quality, and
expect to accelerate the adaptive parameter adjustment
through the convergence speed of the network, but BP
neural network with adaptive parameters is difficult to
determine, it is easy to fall into step large or very small,
also have to be improved in the future.
BP neural network is a nonlinear ability, high precision,
can realize the automatic modification of parameters, and
for the global optimal solution for the, good
approximation ability of the algorithm, has been widely
used in various types of water quality prediction and
warning system in. And BP neural network is a feed
forward neural network, and the common three layer BP
neural network includes input layer, hidden layer and
output layer. Each layer consists of a number of neurons
that can perform addition and multiplication operations,
and the neurons are connected to the next layer of
neurons. BP neural network can be trained by the way of
simulating the excitation and inhibition of the information
of animals, and the effective prediction is completed by
the trained neural network. In order to overcome the shortcomings of traditional
BP neural network, we has presented the adaptive
variable step size algorithm based on fuzzy control theory,
which is used to train a feed-forward Artificial Neural
Network (ANN), reduce the learning time of BP neural
network, improve the convergence efficiency and
network stability. According to the water quality
monitoring data from one Penaeus vannamei breeding
base in Hangzhou, establish a prediction and evaluation
on water quality of the multilayer feedforward network
model. At the time of training, a number of water quality
data at the same time, a number of time points to enter the
neural network, to a number of water quality data for the
tutor to learn. Using the neural network to find the water
quality parameters of the general rules of change in time,
so as to achieve the goal of a number of water quality
parameters at the same time. The results show that the
improved adaptive variable step size BP neural network
by fuzzy method have the characteristics of fast
Journal of Advanced Agricultural Technologies Vol. 3, No. 4, December 2016
©2016 Journal of Advanced Agricultural Technologies 270doi: 10.18178/joaat.3.4.270-275
convergence speed, high prediction accuracy, good
stability and so on, provides a new method for
aquaculture water quality prediction and evaluation.
II. MATERIALS AND METHODS
A. Experimental Materials
The experiment was conducted at the one of breeding
base of the Hangzhou Beijiping Fisheries Science and
Technology Co., Ltd. by the experimental instrument of
multi-parameter water quality analyzer (model: U-52)
made in Japan. The measured data of water quality
parameters was obtained from September 12, 2014 to
September 25, last 12 days. We select the temperature (T),
pH value, Dissolved Oxygen (DO), and the Redox
Potential (ORP) in the parameters as the experimental
data of four indexes. In order to improve the efficiency of
the operation, the twelve day of the 14400 sets of data
(sampling period of 1 minutes) were transferred to the
219 sets of valid data (sampling period of 1 hours) by the
same-interval sampling and the elimination of false data,
We use the 135 groups of data as the training data, the
groups as the test data.
B. BP Neural Network of Aquaculture Water Quality
�
�
Dn-m
Dn-m+1
Dn
wij wjk
i
j
kTn+1
DOn+1
ORPn+1
pHn+1
Dn = Tn DOn ORPnpHn
input layer
hidden layer
output layer
Figure 1. Topological structure of BP-neural network for aquaculture water quality prediction
Temperature (T), pH value, Dissolved Oxygen (DO),
Redox Potential (ORP), is an important parameter to
characterize the quality of aquaculture water. The same
time temperature (T), pH, Dissolved Oxygen (DO),
Redox Potential (ORP) was used as a unit of data in Fig.
1, denoted as D (n). In the actual production, various
water quality parameters is changed gradually, the data at
a certain moment and its historical data are closely related.
In order to make the neural network can be forecasted
according to the historical data of water quality
parameters, we selected the continuous n = 10 sets of
historical data as input, following a time data as output,
neural network were prediction function expression:
( 1) ( ( 1), ( 2), ..., ( ))D t F D t D t D t k
where F(x) expresses the predictive mapping from the
water quality measured data to the water quality
predictive data generasted from the neural network. D(T)
expresses the water quality parameter group in the t time.
The number of nodes in the hidden layer determines
the fitting ability of neural network. In general, the
number of nodes in the hidden layer is more, the fitting
ability of the neural network is stronger, but it is more
difficult to learn and train at the same time. Hidden layer
node number is not yet mature and unified method, but
the number of hidden layer nodes for the general problem
can be determined by empirical formula [6]:
L b c a
where L express the number of hidden layer nodes, b is
the input layer nodes, c is the output layer nodes, a is one
empirical constant. Through experiment and comparison,
we selected b=40, c=4, the number of hidden layer nodes
is 14.
In order to ensure that the neural network has enough
input sensitivity and good fit, reduce adverse effects due
to different numbers of magnitude, do the following
normalization, all data will be unified mapping in [0, 1]:
min
max
'x x
xx x
where xmax is the maximum value in the data, xmin express
the minimum value in the data.
Because the change of water quality data is a gradual
process, so the neural network can be used to find the
corresponding relationship in water quality. In order to
shorten the neural network learning time, we adopt the
time serial input mode shown in Fig. 2, namely neural
network first to N consecutive units of data for input data,
to follow the following the N + 1 data as a tutor for a
learning. After the end of the study, a window moves to
the right of the neural network, the N+2 data as a mentor
for the next study. Whenever the window to pass all the
learning data, one study process is over. According to
experience, it is necessary that over ten thousand times
effective learning to complete a successful study.
D(t+1) D(t+2)D(t+3) D(t+4) D(t+5)D(t+6) D(t+7) D(t+8) D(t+9) D(t+10) D(t+11)D(t+12) �
Past water quality data Going to use water quality dataAre using water quality data
D(t)�
Figure 2. The learning process of neural network
C. The Improved Adaptive Variable Step Size BP Neural
Network Based on Fuzzy Method
Traditional BP neural network needs long training time,
a relatively complex network may be trained more than
ten hours or a few days. In order to solve this problem,
many scholars have proposed the improved the BP neural
network, which is a kind of effective method to change
the learning step size of the neural network to the
changeable learning step. Literature [6], combined with
the method of momentum and variable learning rate
Journal of Advanced Agricultural Technologies Vol. 3, No. 4, December 2016
©2016 Journal of Advanced Agricultural Technologies 271
method [7], with each learning error rate of change of
learning step size is adjusted, is a kind of effective
method, but in to solve the problem of water is used in
the serial input, each calculation error is a local error,
which resulted in the serious error value fluctuations. The
fluctuation of the error is likely to cause the abnormal
change of the step size, which can affect the stability of
the whole study.
Fuzzy method [8] is based on fuzzy mathematics, the
thinking process is absolute, so as to achieve the purpose
of precise and strict. The model using fuzzy algorithm in
the implementation of the decision of stability and
scientific characteristics, using fuzzy algorithm to adjust
the BP neural network to learn the step size and achieve
the value of stable optimization size. The convergence of
BP neural network is more rapid.
Fuzzy algorithm to every neural network learning error
signal generated based on neural network using a set of
fuzzy algorithm to the convergence of judge, and for the
modification of the neural network weights and biases,
malignant learning (is not conducive to weight
convergence of learning) cancellation provides a control
signal.
The specific working flow of the fuzzy controller is
shown in Fig. 3.
normalization processing
Fuzzingfuzzy
decision
control signal output
rule bank
Error signal
input
Figure 3. The working process of the fuzzy controller
In order to effectively adjust the step size, it is
necessary to make a judgment on the state of convergence
by the fuzzy method. The signal of the error change rate
is processed by the rectangular triangular membership
function shown in Fig. 4 into {-3, -2, -1, 0, 1, 2, 3} seven
signal, respectively, on behalf of error high rise, error
medium rise, error low speed rise, error stability, error
low decline, error medium speed decline, error high
decline.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
membership
X(n)
{-3} {-2} {-1} {0} {1} {2} {3}
Figure 4. The membership function
III. RESULTS AND DISCUSSION
A. Results and Analysis of FABPM
The BP neural network was trained with the same set
of water quality data (September 12-20, 2014 at the one
of breeding base of the Hangzhou Beijiping Fisheries
Science and Technology Co., Ltd.) by using the
traditional BP neural network learning algorithm, the
Improved adaptive variable step size BP neural network
based on momentum method [6], and the Improved
adaptive variable step size BP neural network based on
fuzzy method that we proposed in this paper, respectively.
The concrete results as shown in Table I.
TABLE I. THE TRAINING RATE COMPARISON OF SEVERAL ALGORITHM
Conver. mode Target error Step Time Remarks
Traditional BP neural network 0.04%
0.6 Infinite Step size is too large, caught in a shock
0.4 Infinite Step size is too large, caught in a shock
0.2 4.5h Can achieve the goal
0.05 >5h Step size is too small, convergence is too slow
Improved BP neural network based
on momentum method[6] 0.04% variable step size Infinite
The error shock is easy to cause the step size is
too large, can not convergence
Improved BP neural network based
on fuzzy method 0.04%
adaptive variable
step size 3.4h
Can effectively eliminate the error shock, fast
convergence
When the BP neural network was trained by the
traditional BP neural network algorithm, there would be
caught in a shock and not be convergence due to large
step size, such as step is 0.6 and 0.4. In the other case,
such as convergence step is 0.05, due to the step is too
small and a convergence slow phenomenon will be
appeared. As you can see, the choice of step size in the
traditional learning algorithm will be very obvious
influence on the learning effect, thus it is necessary to use
variable step size neural network learning algorithm to be
optimized.
Improvement adaptive variable step size neural
network learning algorithm based on the momentum
method [6] often exhibit unstable phenomenon in the
process of water quality forecast, because there is no
effective treatment of the error in the data, resulting in BP
network not be able to continue to learn.
Presented in this paper, the improved adaptive variable
step size BP neural network learning algorithm with the
fuzzy method can eliminates the error shock impact of
step size, so that the network training adaptive variable
step better and fast convergence.
Journal of Advanced Agricultural Technologies Vol. 3, No. 4, December 2016
©2016 Journal of Advanced Agricultural Technologies 272
B. Results and Analysis of Water Quality Prediction
Model
Effect of neural network training can be used two
indicators to measure, one is the learning effect,
embodies the neural network to reproduce the known
training data, another is prediction effect and
representation of the neural network for the prediction of
the unknown data. Generally speaking, in the case of no
excessive fitting, the better effect of training the neural
network, the prediction effect is better. However, if the
weight or the number of training is not reasonable caused
by the phenomenon of over fitting, but will reduce the
accuracy of prediction.
Fig. 5 is the learning situation of BP neural network for
135 learning data. The solid line shows the water quality
measured data at the same time, the dotted line represents
the output of the corresponding time points of neural
network. If the two points coincide, BP neural network
can be a time period (60 minutes) early to predict the
water quality. From Fig. 6 we can see that the neural
network for four water quality parameters can achieve a
very good learning, the absolute value of the relative
error is less than 5%. Dissolved Oxygen (DO) in the 70
data appear obvious decline, after verification is because
of thunderstorm weather caused, even so, neural network
also achieved good prediction, which also from another
aspect illustrates the neural network has a good ability to
learn.
Figure 5. Outcomes of model training
Fig. 6 is the remaining 85 test data into the trained
neural network to obtain the prediction data, in which the
solid line represent water of the real data and the dotted
line represents the data for the prediction of the water
quality. If the two coincidence that BP neural network
can advance a period of time (60 minutes) to predict the
change of water quality.
It can be seen in the course of the breeding pH, DO
fluctuations are relatively small, the temperature and
ORP fluctuations are more obvious and the neural
network can be an effective forecast. Temperature (T),
Dissolved Oxygen (DO), the maximum error of pH value
is less than 2%, ORP due to the change is more severe
and the maximum error is less than 10%.
Journal of Advanced Agricultural Technologies Vol. 3, No. 4, December 2016
©2016 Journal of Advanced Agricultural Technologies 273
Figure 6. Outcomes of model forecast
Significance of dissolved oxygen (do) is a variation of
relatively rapid parameters for breeding, the dissolved
oxygen content is too low will lead to floating aquatic
products or choked to death. From the graph we can see,
the dissolved oxygen prediction value and the
corresponding actual values coincide, we assume that
when dissolved oxygen content lower than 3.45, the need
for alarm. According to the calculation of the BP neural
network model for water quality prediction, dissolved
oxygen in the 52nd sampling period reach the alarm value,
than the actual early warning for 60 minutes, farming
enterprises can use the 60 minutes, take corresponding
measures, reduced the economic loss of the
corresponding.
Journal of Advanced Agricultural Technologies Vol. 3, No. 4, December 2016
©2016 Journal of Advanced Agricultural Technologies 274
IV. CONCLUSIONS
In order to overcome the slow convergence speed, easy
to fall into the shock and generalization ability is not
strong and so on of traditional BP neural network, the
adaptive variable step size BP neural network learning
algorithm based on fuzzy method were developed, which
can reduce the learning time of BP neural network,
improve the convergence efficiency and network stability.
According to the water quality monitoring data from one
of Penaeus vannamei breeding base in Hangzhou, a
mathematical model of multi-layer feed forward network
was established to predict and evaluate the quality of the
aquaculture water. At the time of training, a number of
water quality data at the same time, a number of time
points to enter the neural network, to a number of water
quality data for the tutor to learn. Using the neural
network to find the water quality parameters in the time
of the general rule, so as to achieve the output of a
number of water quality parameters of the target. The
results show that the improved adaptive variable step size
BP neural network with fuzzy method has the
characteristics of fast convergence, high accuracy and
good stability. It provides a new method for water quality
prediction and evaluation of aquaculture.
ACKNOWLEDGMENT
This work was supported in part by a grant from
Zhejiang Province major science and technology projects
(2012C03009-2) and Hangzhou science and technology
development plan (20140432B08).
REFERENCES
[1] C. Ma, D. Zhao, J. Wang, Y. Chen, and Y. Li, “Intelligent
monitoring system for aquaculture dissolved oxygen in pond
based on wireless sensor network,” Transactions of the Chinese
Society of Agricultural Engineering, vol. 31, no. 7, pp. 193-200,
2015.
[2] L. Zhang, et al., “Monitoring of water quality in aquaculture based
wireless sensor networks,” Computer Knowledge and Technology,
vol. 5, 2015.
[3] H. Pan, J. Guan, Y. Shi, and T. Li, “Design and test of aquaculture
water environment monitoring system based on w ireless sensor
network,” Journal of Chinese Agricultural Mechanization, vol. 5,
pp. 246-250, 2014.
[4] S. Liu, et al., “Forecasting model for pH value of aquaculture
water quality based on PCA-MCAFA-LSSVM,” Transactions of
the Chinese Society for Agricultural Machinery, vol. 5, pp.239-
246, 2014.
[5] H. Hu, et al., “LM - BP neural network application in water
quality forecast,” Microcomputer Applications, vol. 9, pp. 44-46,
2011.
[6] Q. Yuan, J. Huang, X. Fu, and S. Weng, “Study on prediction
model of aquaculture water quality based on artificial neural
network,” Hubei Agricultural Sciences, vol. 1, pp. 143-146, 2013.
[7] J. Liu, K. Han, and H. Li, “Application of an improved arithmetic
on BP neural networks,” Journal of North China Institute of
Technology, vol. 6, pp. 449-451, 2002.
[8] L. Lin, T. Feng, W. Zhou, and H. Zheng, “Expert system based on
neuro-fuzzy approach for sensory quality assessment of tobaccos,”
Journal of Ocean University of China (Natural Science), vol. 6, pp.
931-936, 2001.
Ding Jin-ting was born in Hangzhou, Zhejiang, China, on March 21st, 1968. Bachelor of Automation given at Tianjin, China, in July 1997,
graduated from the Tianjin University. Master of Electrical Engineering
Power Electronics and Drives given at Hangzhou, China, on September 30th, 2006, graduated from the Zhejiang University. She worked in
Zhejiang University City College of information and electrical engineering school, interesting in the intelligence of agriculture,
agricultural information, also interesting in control engineering and
automation.
Journal of Advanced Agricultural Technologies Vol. 3, No. 4, December 2016
©2016 Journal of Advanced Agricultural Technologies 275